Brain tissue that generates its own electricity.

A brain organoid begins as a small population of human stem cells — often reprogrammed from an ordinary blood or skin sample — guided through the same chemical signaling steps that direct a developing brain.

Over the course of several weeks, those cells differentiate into neurons and supporting glial cells, and self-organize into a three-dimensional structure a few millimetres across, roughly the size of a lentil. There is no skull, no body, no senses, no environment to act in. What there is, reliably and reproducibly, is electrical activity.

That activity does not appear all at once, and it does not stay the same. Early cultures fire sparsely, with little structure. Over weeks to months, recordings begin to show coordinated bursts — many neurons firing together, falling silent, then firing again — a signature of maturing connectivity that has been tracked since some of the field's earliest published recordings. An organoid is not thinking in any sense comparable to a brain embedded in a body with a world to respond to. But it is computing something: a living network turning electrical and chemical input into structured, time-varying output.

To see that output, researchers place the organoid onto a microelectrode array (MEA) — a grid of electrodes listening for the electrical signature of nearby firing neurons. The result is a dense stream of spike data: which electrode, at what time, at what amplitude. Turning that stream into something interpretable — firing rate, synchrony between regions, network bursts, large-scale structure — is itself a substantial computational problem, and one with very little standardization between labs.

In 2023, a research group at Johns Hopkins proposed organoid intelligence (OI) as a field in its own right: using brain organoids not only as disease models, but as a substrate for biological computing, testing whether structured information processing — and eventually learning — can be supported by living neural tissue outside a body. Part of the motivation is practical. A human brain runs on roughly twenty watts, a small fraction of what comparable artificial systems require, and biological tissue offers a fundamentally different kind of computational substrate to study and compare against silicon. Part of the motivation is scientific: organoids let researchers ask developmental and computational questions about neural tissue that are difficult, or simply impossible, to ask inside a living human brain.

The field is young. Its first widely cited organoid electrophysiology recordings date to 2019, and as of 2026 there is still no standard analysis pipeline, no agreed file format, and very little publicly available data to work with. Five projects, presented below, are an attempt to work inside that gap — building the tools the field is missing, characterizing what real recordings actually look like over time, identifying exactly where the standard recording method falls short, and naming an ethical question the field has not yet formulated directly.

Reading, shaping, and questioning the signal.

A reflex borrowed from every visually-capable animal.

A rapidly expanding shape — the visual signature of something approaching — triggers an immediate, unlearned defensive response across insects, fish, rodents, and primates alike: the looming stimulus. loom-oi drives a simulated organoid's electrodes directly from a clip's real, computed expansion rate, and trains a reward-calibrated readout to call it calm or threat, starting blind.

Calibrating it surfaced a genuine failure: the "calm" control condition's own measurement noise exceeded a real, weak threat signal — found only by checking per-class accuracy rather than an aggregate score, and fixed at its source. After correction, 5 of 15 random seeds pass a three-stage qualification screen, and under an equal real-time budget the organoid readout reaches 86% against a standard digital model's 100% — a genuine, informative comparison, not the degenerate 100%/100% an under-calibrated version produced.

A working demonstration, verified end to end.

organoid-oi's current version demonstrates a simple, honestly-named mechanism working reliably: a simulated organoid's own firing intensity scales directly with how strongly it is driven — a longer input produces more firing, a physical relationship that requires no training. What is trained, through Three-Factor spike-timing-dependent plasticity and a local reward-calibrated readout layer, is the ability to read that signal correctly, starting from a blind, uncalibrated state.

Applied to word length, the result is a system that takes a set of scrambled words and correctly assembles them into a sentence, and correctly places a newly introduced word into an existing one. We call this a reflexive demonstration, deliberately — the same word used by Kagan et al. for their organoid's closed-loop Pong behavior in 2022. The mechanism is simple and named plainly; the result is real.

Three things were verified before this was called done. The dose-response relationship itself — word length against firing intensity, correlation +0.999 across eight test words, charted below. Readout calibration — accuracy rising from near-chance to reliable correct ordering as reward and penalty shape the decoder. And interface parity with real hardware: the system is built against a single method signature, implemented identically for simulation and for a live-tissue backend, so the same code runs on either — pending the API integration step that connects the second one.

This version followed an earlier, rigorously negative attempt at a harder claim — categorical rather than reflexive learning, tested across several task designs and difficulty levels, none of which held up under repeated verification. That result was published in full rather than quietly reworked, and it directly motivated the smaller, verified claim reported here. The six measurement defects found along the way — and the general failure pattern they share — are documented separately as a standalone methodological contribution: doi.org/10.5281/zenodo.21892103.

Functional provenance stratification.

The organoid intelligence ethics literature has developed along three largely separate axes: consciousness and moral status, informed consent and the right to withdraw, and the commodification of donated biological material. Each is reasonably well documented on its own. None directly names the following scenario.

This is not purely speculative. Large-scale iPSC biobanking already catalogues donor lines by differentiation efficiency and functional yield — an organoid-intelligence-specific performance axis would extend an existing practice, not invent a new one. The scenario links two risks not previously connected in this combination: retrospective consent-scope violation, when a line is reclassified by a criterion that did not exist at the time of donation, and structural inequity, when sampling bias causes specific donor populations to be disproportionately coded as lower-performing.

I term this practice functional provenance stratification and propose four minimal safeguards: explicit disclosure in consent forms, transparency and auditability of performance metrics, periodic equity analysis across donor populations, and preserved withdrawal rights independent of a line's performance category. This is a conceptual contribution, offered in the spirit of the anticipatory neuroethics already established for questions of organoid consciousness — not an empirical finding.

Independent, by necessity and by choice.

I work without a laboratory and without formal academic affiliation — using public datasets (DANDI Archive) and direct data access granted by research platforms. Based in Aliağa, İzmir, Turkey.